Apollo Global Management and Blackstone closed a $35 billion debt financing package that grants Anthropic structured access to Alphabet's proprietary tensor processing units through a special purpose vehicle. The arrangement, which surfaces weeks after Anthropic disclosed its Claude 3.5 training costs exceeded internal projections by 40%, represents the largest credit facility ever extended against leased computational infrastructure rather than physical assets.
The SPV holds no silicon. It holds contract rights. Anthropic draws compute capacity from Alphabet's custom chip inventory under a take-or-pay agreement running through 2029, with minimum annual utilization floors pegged to 18 exaflops of training throughput. Apollo and Blackstone priced the senior tranche at SOFR plus 285 basis points, a spread that implies lenders view access to frontier chips as investment-grade collateral in a market where NVIDIA H100 lease rates now exceed the per-unit purchase price every eleven months. The subordinated tranche, unrated and held entirely by Blackstone's opportunistic credit vehicles, carries a spread north of 650 basis points.
This matters because credit markets just validated compute-as-a-service infrastructure at scale, without requiring model IP as security. Anthropic pledged no equity. It pledged no algorithms. The collateral package consists of Alphabet chip capacity, Anthropic's investment-grade treasury management, and a $4.2 billion equity backstop from the existing investor consortium led by Menlo Ventures and Spark Capital. Lenders accepted exposure to a single counterparty chipmaker because Alphabet's TPU v5 architecture already powers 60% of Google's internal inference workloads, creating natural refinancing demand if Anthropic defaults. The debt matures in tranches across six years, with covenants tied to minimum model benchmark scores rather than revenue milestones.
The structure signals three shifts. First, alternative credit managers now underwrite AI infrastructure at sovereign-debt scale, treating compute access like toll-road concessions or telecom spectrum. Second, Alphabet monetized its captive silicon R&D without selling chips outright, sidestepping export-control scrutiny while locking in a $35 billion demand commitment that justifies expanding TPU fab capacity in Oregon and Taiwan. Third, Anthropic secured frontier compute without further equity dilution at a moment when its whisper valuation sits near $60 billion, up from $18 billion fourteen months ago. The SPV lets Anthropic report chip expenses as operating leases rather than capital expenditures, improving reported free cash flow by an estimated $320 million annually through 2027.
Operators should track three follow-on events. Alphabet reports Q1 earnings April 24, where management will face questions on whether this structure becomes a template for Meta, xAI, or Microsoft-backed OpenAI. The Federal Reserve's May meeting will test whether leveraged debt tied to compute infrastructure counts toward systemically important nonbank exposure, given Apollo's existing $87 billion in AI-adjacent credit commitments. And Anthropic's next model release, expected mid-June, will clarify whether the SPV's utilization floors reflect actual training cadence or speculative capacity hoarding.
Blackstone's infrastructure group now holds senior positions in $58 billion of AI compute financing closed since November 2023, more than double its total energy-transition debt book.